I simultaneously think that AI (as in LLMs) is probably a net negative socially, and that many, many high profile professional critics of it don't have a fucking clue what they're talking about. - ThreadSky | a Reddit-style client for Bluesky

opinionhaver.bsky.social • 35 days ago

I simultaneously think that AI (as in LLMs) is probably a net negative socially, and that many, many high profile professional critics of it don't have a fucking clue what they're talking about.

Reposted from rev. howard arson

like one of the big things about 404 media, Brian Merchant, Paris Marx, Ed Zitron, etc., is that they neither know nor care how the subject of their criticism actually works

Comments

digifox.binaryden.net•35 days ago

I think it'll be a net positive once the fad dies down

sagastar.bsky.social•35 days ago

if the proponents don't know, why should the critics?

thebadcode.com•35 days ago

It’s really unfortunate that the LLM boom came right on the heels of Crypto Madness, which was, in fact, a huge scam that never could have done what its proponents claimed. Now all these guys like Ed feel comfortable saying “well, I called it correctly last time”

terrycrows.bsky.social•35 days ago

Not helping is the fact that the same VC hype men from the crypto scene seamlessly became AI boosters.

opinionhaver.bsky.social•35 days ago

LLMs are potentially enormous time savers if you do things with code or technical writing. this is a huge portion of the work that people in SWE type roles do, and it is a much better explanation for why silicon valley is so taken with the tech rather than elaborate conspiracy theories.

proudanselmo.bsky.social•35 days ago

They're also great when you're learning to code, because they can tell you quite easily how to code for X, but with this other Y variable, except in cases of Z, instead of you having to read through incredibly dense technical manuals to try and figure out how all three fit together.

boomies-n-zoomies.bsky.social•35 days ago

Eh, I’m really not convinced they’re very good at coding tbh.

proudanselmo.bsky.social•34 days ago

Why not?

boomies-n-zoomies.bsky.social•34 days ago

Every time I’ve seen someone use them, they tend to produce a lot of subtle bugs that are net negative in the long run. And they basically cannot write code that has a higher correctness requirement than say, Java.

hamburglarhimbo.bsky.social•35 days ago

I work in SWE and I'm reticent to use it for coding because I don't want client code leaked to the public.

hdevalence.bsky.social•35 days ago

so pay for a service that doesn’t do that?

hamburglarhimbo.bsky.social•34 days ago

ChatGPT Enterprise is the only plan that lets you exclude your data from training and control retention windows, and also has no published price tag. Probably not worth the cost at the individual level.

hdevalence.bsky.social•34 days ago

use claude sonnet with cursor

rocketpilot.xyz•35 days ago

And they legally indemnify you against accidental leakage?

hdevalence.bsky.social•35 days ago

literally no one cares about this. get real

rocketpilot.xyz•35 days ago

Wow that extremely puts my mind at ease! Thank you!

botsnack.bsky.social•35 days ago

Feels like there's a profound fallacy embedded in the idea of "well this works very well in this highly prescriptive state machine, should replace humans soon."

Programming is weird. I've been reminded of that every one of the 5 times I've tried getting into it, since I was 8yo.

proudanselmo.bsky.social•34 days ago

There's a tendency by the techbros of the time to believe that the development of a significant new technology will revolutionize humanity and take over all our menial jobs, etc., and each time, the techbros overpromise and underdeliver.

I don't see why AI will be any different.

weirdhimejin.bsky.social•35 days ago

Sure you can use LLMs for code generation, but we programmers have had code generation for decades now. LLMs are the most expensive way to do the job

botsnack.bsky.social•35 days ago

What was before LLMs?
Only familiar with basic auto-linting/completions/etc

weirdhimejin.bsky.social•35 days ago

Before I was born it was just called Meta-programming. The first mainstream version I can think of is Ruby on Rails's Active Record

dollspace.gay•35 days ago

And AI image generators are *incredibily* popular among common people. Sure people call it slop, but the median citizen loves slop and they always have. Thomas Kinkade was popular for a reason.

blank.org•35 days ago

Also: speed. Speed speed speed. Film replaced paint, CCDs replaced film, speed rules everything around me. Yes, you can commission an artist, but not in under five minutes. If the image is just going into a PowerPoint deck that will be seen ten times and forgotten? Speed is everything.

dollspace.gay•35 days ago

An "ok enough" image for a powerpoint made in 30 second is going to win out over paying someone to make it in three days

blank.org•35 days ago

Exactly. And ditto for an email or a text message you’re sending to your family.

giziti.bsky.social•35 days ago

nice for converting unstructured input -> structured output

opinionhaver.bsky.social•35 days ago

I do sort of feel like being a quant social science academic gives you a unique perspective on this stuff because you're highly exposed to both the upsides and downsides. it's hugely helpful for cleaning datasets and coding. it's a disaster for teaching.

dawsonvosburg.bsky.social•35 days ago

as a qual social scientist it is good for transcribing and other stuff that is 1. basically tedium and 2. based on preexisting human work that supplies most of the actual core material. not super good at stuff that is at its core analytical and I wouldn’t even really want it to do that for me

ondankbare-plant.bsky.social•35 days ago

In biomedical research, we already had a huge problem with made up BS from Chinese tadpole paper mills, but with AI dross it's become even easier, save for the images of giant rat penises and nonsensical signaling diagrams. But, at the same time it's transformative for digital pathology.

ondankbare-plant.bsky.social•35 days ago

For the uninitiated: actual published figures from a frontiers journal article last year, reportedly "peer reviewed". Has been retracted after well, just look at this shit.

Comment image

machminotaur.bsky.social•35 days ago

the AI desires to give the rats HUGE HOGS

machminotaur.bsky.social•35 days ago

wait. I think it's doing that because it can't discriminate between input sources. it has a thousand petaflops or whatever of porn and goes "this shape is always huge" so it makes it disproportionally huge because of all the porn.

ondankbare-plant.bsky.social•35 days ago

It was quite exhilarating to be able to talk about giant rat penises in a professional context, when AI integration into the workplace was discussed.

aaronfernandez.bsky.social•35 days ago

100% my experience as well